Motivation
A reflecting zero means a new arrival can restart a population. An absorbing zero means extinction is permanent. Reusing a reflecting-boundary stationary formula would answer the wrong question: the point mass at zero is now stationary, but it says nothing about whether extinction occurs from a positive initial population or how long it takes.
Two kinds of limits need separate justification. Moving an upper boundary outward gives eventual extinction probabilities. Taking expected passage times through that limit also requires control of the process lifetime and a nonnegative convergence argument. A formal solution of an infinite difference equation is not enough. We therefore solve the finite boundary problem first and preserve its probabilistic meaning as the boundary tends to infinity.
The preceding non-explosion and renewal notes supply the proper-function criterion and basic nonnegative convergence tools.
Definitions
State zero is absorbing: . For every , assume finite rates and to neighbors and . The minimal chain is assumed non-explosive. A useful sufficient condition is with arbitrary finite death rates: is proper and at positive states, while . Here denotes the continuous-time generator.
Starting from , write , allowing infinity. Extinction means .
Definition 3.4.1
(Scale increments, scale sums and speed weights)
For the absorbing chain above, define
The last identity follows by canceling the finite products. These weights are anchored at one; dividing by a birth rate at zero would be invalid because that rate is zero. A pure-death model, with births zero, is treated separately as an elementary boundary case rather than by products containing division by zero.
Definition 3.4.2
(Finite-boundary time Green function)
For an integer upper boundary , let . The expected occupation before either boundary is
This is a time Green function: it measures time spent at a state, not just the number of visits to it.
Theorem/proposition
Theorem 3.4.3
(Scale and finite-boundary Green formulas)
For and ,
Both quantities arise from a finite absorbing problem with finite mean termination time. No infinite-boundary condition is inserted into these formulas.
Theorem 3.4.4
(Extinction and mean-time series)
Let . Under the stated non-explosion assumption,
If extinction is certain, then, as an equality in ,
Thus, in the certain-extinction case, the mean is finite for one, hence every, positive finite starting state exactly when . If extinction has probability below one, the unconditional mean is infinite.
Theorem 3.4.5
(Linear branching extinction law)
For , , with , , let . Then
Extinction is certain for , and has probability for . For one ancestor and ,
At criticality the mean is infinite. Above criticality the unconditional mean is infinite; conditional on eventual extinction, the one-ancestor mean is .
Proof sketch or proof idea
Solve the finite boundary problem
Proof
(Scale and Green formulas on a finite interval)
Before reaching zero or , only finitely many states occur. From each there is a positive probability of following a path to a boundary within a common finite number of embedded steps. A geometric survival bound gives a finite expected number of steps. The positive departure rates on this finite interior have a positive minimum, so mean holding times are uniformly bounded. Consequently .
First-jump conditioning for gives . Therefore is proportional to . Summing differences and imposing , proves the scale formula.
Fix . The occupation column satisfies on the interior and is zero on both boundaries. Indeed one visit to initial state contributes mean time if , followed by the destination's expected continuation. Away from , the displayed Green candidate is linear in the scale on each side, hence harmonic. At its two scale slopes change. Using and , their jump gives exactly . Both boundary values vanish. Uniqueness follows from the finite absorbing boundary problem: a harmonic difference with zero boundary must have zero expectation at termination, or equivalently obey the finite maximum principle. Thus the candidate is the actual expected occupation.
Remove the upper boundary carefully
Proof
(Monotone limits of the finite-boundary quantities)
The events decrease as increases. A path avoiding zero cannot remain forever in a fixed finite interval: the finite-region geometric argument rules that out. Hence their intersection is the never-extinct event up to a null set. Letting increase in the scale formula proves the extinction probability.
The times increase to , interpreted as infinity when extinction fails. On an extinction path, the finite history before hitting zero has a maximum state, so for all sufficiently large . On a never-extinct path, non-explosion ensures that the upper-boundary hitting times tend to infinity rather than accumulate at a finite lifetime. This is the role of the lifetime assumption; in the certain-extinction case, eventual equality already justifies the stopping-time limit. For a fixed , monotone convergence of the occupation integrals gives the infinite-boundary Green function. If , its value is
Sum over all and use Tonelli to exchange nonnegative sums. A term occurs exactly when and . Therefore
No finiteness was assumed in this passage. The first outer term already contains the entire positive speed-weight sum, while only finitely many positive outer coefficients occur. This proves the finiteness criterion. If extinction fails with positive probability, an infinite value of has positive mass, forcing its unconditional expectation to infinity.
Proof X-ray
Why the mean series is trustworthy
The argument has three layers, and each hypothesis is used in one of them. Finiteness of the interior and positivity of its rates give ; this licenses first-jump conditioning and the uniqueness of the bounded solution, and it needs no assumption at infinity.
The key move is the jump condition at . The identity is exactly what makes the change of slope of the Green candidate produce , so the speed weights are forced by the generator rather than chosen.
Passing to uses non-explosion: on paths that escape upward it makes increase to infinity rather than to a finite explosion time, so the limit measures time before extinction. Monotone convergence then carries the occupation integrals to the limit while allowing the value , and only after that is the nonnegative double sum rearranged by Tonelli. The formula therefore never presupposes that the mean is finite.
Build the branching law, not just a drift equation
Proof
(First-event equation for one family)
Give each individual independent birth and death clocks of rates and . A birth creates one additional individual, and a death removes one. From population the total rates are and . The proper-function criterion above proves non-explosion, and independent descendant families give the same chain law by uniqueness of its finite jump construction. With original individuals, extinction by time occurs when all families have become extinct, giving .
For one ancestor, condition on its first event. A death at time leaves no descendants. A birth leaves two independent families, the newborn and the original individual's fresh future. Exponential memorylessness justifies that fresh future. Thus
Changing the integration variable and differentiating gives . The right side is locally Lipschitz, so the initial-value solution is unique; direct separation or substitution gives the displayed formula. Its limit gives all three extinction regimes. At it solves , yielding .
For , put . The one-family survival probability is . Integrating its tail and substituting gives
At criticality the survival tail is , whose integral diverges; from ancestors survival is at least this large. For , positive nonextinction probability makes the unconditional mean infinite.
The conditional statement also needs a derivation. Since extinction by time implies eventual extinction, for ,
This is exactly the subcritical extinction-time distribution with the parameters interchanged. Integrating the one-ancestor tail by the calculation just completed gives . No assumption that conditioning “obviously swaps the rates” is used; the full conditional distribution was derived before its mean.
Proof X-ray
Where the quadratic term comes from
Two hypotheses carry the argument. Independent individual clocks give the branching property: separate ancestors found independent families, so the law from ancestors is the -th power of the one-family law. Memorylessness lets a birth be read as a restart with two fresh, independent families, which is why the first-event equation contains , while a death contributes the constant . Differentiation turns the integral equation into the Riccati equation , and local Lipschitz continuity makes select one solution. The conditional law needs no new probability: dividing by closes the argument.
The branching law separates two pieces of information that the general formulas also keep apart: where the chain ends and how long it takes.
Concept lensStructural
An extinction probability has no clock unit
Multiplying every positive-state birth and death rate by the same constant leaves their ratios unchanged. The scale increments and eventual extinction probabilities are therefore unchanged, while every holding time is divided by that constant. In the mean-time series, the speed weights stay the same and each denominator containing a death rate acquires the same factor. Thus the mean passage time is divided by the factor whenever finite, and remains infinite otherwise. This comparison is a useful dimensional and conceptual check: information about which boundary is reached cannot determine the time needed to reach it. Even when extinction is certain in both models, the clocks can run at different speeds. The explicit branching distribution records both kinds of information; its infinite-time limit retains only the extinction probability.
Worked examples
Worked example 3.4.6
(Constant positive-state rates)
Let births have rate one and deaths rate two whenever the population is positive, with zero absorbing. Rates are bounded, so non-explosion is immediate. The scale increments are , so extinction is certain. Speed weights are , and the mean formula gives . From one, the chance of hitting three before zero is . More generally constant rates give mean ; equality gives certain extinction with infinite mean, while gives extinction probability .
Worked example 3.4.7
(Two ancestors do not double the extinction time)
Take linear branching with . Then and . The mean from one is . From two, the mean-time series adds a second term: and . Hence
about , rather than . Descendant families evolve concurrently; extinction of both is the maximum of their independent family extinction times, not their sum. The first-moment population drift alone would not supply this exact passage-time calculation.
Worked example 3.4.8
(Critical and conditional supercritical boundaries)
For , extinction by time from one has probability . It tends to one, but its survival tail has infinite integral. Certain extinction does not imply finite mean time. For , eventual extinction has probability from one. The unconditional mean is infinite, while the conditional extinction-time law equals the subcritical law, with mean . These are different expectations on different experiments, not contradictory answers.
Common mistakes
Common mistake
(Transplanting the reflecting-boundary formula)
An absorbing zero changes the boundary equation. The reflecting model's zero-to-zero cycle formula describes returns to zero followed by a restart; with zero absorbing there is no restart, so that formula answers no question about extinction.
Common mistake
(Trusting a formal solution of the mean equation)
The equation on positive states with does not determine the mean. For constant rates , , the function solves it for every , because and . Every such is nonnegative, but only is the expected passage time; the Green limit selects it, and retains infinity where appropriate.
Common mistake
(Ignoring the never-extinct event in a mean)
Above criticality, with positive probability, so the unconditional mean is infinite. A finite answer such as is a conditional mean; it requires naming the conditioning event and deriving the conditional law first.
A stationary distribution concentrated at the absorbing state is a further trap, since it says nothing about whether the chain reaches that state.
Counterexample mode
A stationary point mass does not force extinction
Claim: since the point mass at zero is stationary, extinction is almost sure from every state. Take linear branching with , . Zero is absorbing, so its generator row vanishes and the point mass at zero is stationary. Yet , so and , and the extinction formula gives , in agreement with . The correct criterion is the scale series: extinction is certain from every state exactly when .
Replacing an infinite upper boundary by a finite absorbing one generally shortens the experiment; replacing it by a reflecting one creates a different model again.
Summary
Scale differences determine finite-boundary hitting and extinction probabilities. Finite Green functions and non-explosive boundary limits justify the mean-time series without assuming it is finite. Linear branching permits an explicit first-event equation, yielding certain extinction below and at criticality, positive survival above it, and a separately derived conditional extinction-time law.
Exercises
Quick check
(With zero absorbing and constant positive-state rates , start at two. Find extinction probability, mean extinction time and the chance of hitting four before zero.)
Compute and use the constant-rate example.
Solution · Answer
Scale increments have divergent sum, so extinction probability is one. The mean is . Also and , so the finite-boundary probability is .
Quick check
(In a pure-death process with rate from to , start at . Find the extinction-time mean and distribution. Explain why the positive-birth scale products are not used directly.)
Follow the original individuals separately.
Solution
There are successive independent holds of means , giving mean . Equivalently each original individual has independent exponential lifetime of rate , so extinction is their maximum and . Birth rates are zero, so products dividing by them are undefined; this finite-step direct construction supplies the boundary case.
Quick check
(For linear branching starting with three ancestors, find the extinction-time distribution and decide whether its mean is finite.)
Compare the survival probability with that of one family.
Solution · Answer
The distribution is , tending to one. Its survival is at least because for . The tail integral therefore diverges, so the mean is infinite.
Quick check
(For linear branching starting with one ancestor, find eventual extinction probability, unconditional mean and mean conditional on extinction. Derive the conditional distribution used in the last calculation.)
Divide the finite-time extinction probability by the eventual one.
Solution
Extinction probability is , and the unconditional mean is infinite. The conditional CDF is , obtained by dividing the finite-time extinction probability by the eventual one. This equals the subcritical birth-one/death-three CDF; tail integration gives for its mean. The conditioning, not an omitted nonextinction tail, makes that mean finite.